2018-06-04 9:24 GMT+02:00 Heikki Linnakangas <hlinn...@iki.fi>: > On 04/06/18 09:12, Pavel Stehule wrote: > >> 2018-06-04 8:35 GMT+02:00 Simon Riggs <si...@2ndquadrant.com>: >> >>> >>> Sounds good. I think this would need to be restricted by operator and >>> datatype, since in general you won't know if the datatype functions >>> need a snapshot or not. Immutable functions for the operators ought to >>> do it, but I think that might not be enough. >>> >> >> It requires introduction of new "safe" functions (& operators). Immutable >> functions are not enough safe. >> >> CREATE OR REPLACE FUNCTION public.fx() >> RETURNS integer >> LANGUAGE plpgsql >> IMMUTABLE >> AS $function$ >> BEGIN >> RETURN (SELECT count(*) FROM pg_class); >> END; >> $function$ >> >> postgres=# SELECT fx(); >> ┌─────┐ >> │ fx │ >> ╞═════╡ >> │ 343 │ >> └─────┘ >> (1 row) >> > > That function is incorrectly marked as IMMUTABLE. In that situation, it's > enough that we throw a sane error like "ERROR: no snapshot available". >
Yes, it is incorrect mark. Unfortunately - this is often workaround for wrong estimations - so I afraid, in this case, your proposed fix breaks lot of applications. Regards Pavel > - Heikki >